Graphcuts for General Quadratic Binary Problems

We propose a novel approach to optimizing unconstrained quadratic binary problems (QBP) and binary MRFs based on discrete, submodular majorization of the original problem that can be solved efficiently using graph cuts. This yields an efficient algorithm that is appropriate for quickly finding a good solution to general Quadratic Binary Problems. It first solves a majorizing function exactly to get a good global location, then refines the solution using targeted local optimizations. When combined with a tabu-search like approach, it is a competitive technique for solving general QBPs. The technique compares well to several existing techniques.